Abstract
In the field of rock mechanics and engineering, the fusion analysis of multi-source heterogeneous data such as lithology and stress constitutes the core foundation for geological disaster early warning and underground engineering stability assessment. Current rock data modeling based on Graph Convolutional Networks (GCN) exhibits two critical issues: first, features of dominant categories are continuously reinforced through neighborhood propagation, triggering "popularity bias"; second, the inherent over-smoothing problem of GCN leads to blurred local features. To address these issues, this paper proposes a Graph Convolutional framework integrating bias correction, collaborative signal enhancement, and de-smoothing strategies (PCD-GCN). This framework comprises three core modules: the collaborative signal enhancement module quantifies the collaborative contributions of multi-source rock data, strengthening the expression of multi-field coupling information of "lithology-stress-deformation"; the bias correction module combines rock cluster analysis with iterative reweighting strategies to alleviate the excessive dominance of advantaged data; the de-smoothing strategy suppresses global topological influence through vector perturbation, optimizing from the dual dimensions of "global suppression-local preservation". The study validates the effectiveness of PCD-GCN in bias correction and over-smoothing suppression, providing a precise modeling paradigm for geotechnical geological disaster early warning and underground engineering stability analysis, and also offering a novel technical approach for cross-scale correlation analysis of "data-structure-mechanical response" in rock mechanics.
Full Text
Graph Convolutional Model for Geotechnical Multi-source Data Integrating Bias Correction and De-smoothing Strategies
Yu Zhang¹,², Qianhui Ding¹, Siying Wang¹
¹ School of Electrical and Information Engineering & Beijing Municipal Key Laboratory of Urban Architecture Super Intelligent Technology, Beijing University of Civil Engineering and Architecture, Beijing 100044, China
² State Key Laboratory for GeoMechanics and Deep Underground Engineering, China University of Mining & Technology, Beijing 100083, China
Abstract
In the field of rock mechanics and engineering, the fusion analysis of multi-source heterogeneous geotechnical data—including lithology, stress, and deformation measurements—constitutes the fundamental basis for geological hazard early warning and underground engineering stability assessment. However, current rock data modeling approaches based on Graph Convolutional Networks (GCN) suffer from two critical limitations. First, features from dominant categories are continuously reinforced through neighborhood propagation, triggering a "popularity bias" that distorts the learning process. Second, the inherent over-smoothing problem in GCN architectures causes local feature blurring, compromising the model's ability to capture fine-grained geological patterns.
To address these challenges, this paper proposes a novel graph convolutional framework that integrates bias correction, collaborative signal enhancement, and de-smoothing strategies (PCD-GCN). The framework comprises three core modules: (1) a collaborative signal enhancement module that quantifies the cooperative contributions of multi-source geotechnical data to strengthen the representation of multi-field coupling information among lithology, stress, and deformation; (2) a bias correction module that combines rock clustering analysis with iterative reweighting strategies to mitigate the excessive dominance of advantageous data categories; and (3) a de-smoothing strategy that suppresses global topological influence through vector perturbation. These modules work synergistically to optimize the model from dual dimensions of "global suppression and local preservation."
Experimental validation demonstrates the effectiveness of PCD-GCN in both bias correction and over-smoothing suppression. The proposed framework provides a precise modeling paradigm for geotechnical hazard early warning and underground engineering stability analysis, while offering a novel technical approach for cross-scale correlation analysis of "data-structure-mechanical response" in rock mechanics.
Keywords: Graph Convolutional Network; Recommendation Algorithm; Bias Correction; Multi-source Data